Shieldstral

2026-07-28

AI Summary

This work addresses the challenges of large model size and limited policy adaptability in multimodal content safety classification by unifying moderation tasks into a binary question-answering framework and introducing a lightweight, policy-adaptive multimodal classifier. Leveraging a heterogeneous safety data recipe comprising 54.1 million samples and a fine-grained evaluation benchmark, the approach combines a 3-billion-parameter multimodal Transformer with a unified question-answering fine-tuning strategy. The resulting model matches or exceeds the performance of existing models nearly seven times its size on text safety benchmarks and establishes a new state-of-the-art on multimodal safety classification tasks.

Abstract

We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$\times$ its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.

Problem

• multimodal safety classification
• content moderation
• policy-adaptive
• safety classifier
• binary question-answering

Innovation

• policy-adaptive
• multimodal safety classification
• binary question-answering
• heterogeneous dataset consolidation
• content moderation